NBER WORKING PAPER SERIES RE-EXAMINING THE EFFECTS OF TRADING WITH CHINA ON LOCAL LABOR MARKETS: A SUPPLY CHAIN PERSPECTIVE

Transcription

1 NBER WORKING PAPER SERIES RE-EXAMINING THE EFFECTS OF TRADING WITH CHINA ON LOCAL LABOR MARKETS: A SUPPLY CHAIN PERSPECTIVE Zhi Wang Shang-Jin Wei Xinding Yu Kunfu Zhu Working Paper NATIONAL BUREAU OF ECONOMIC RESEARCH 1050 Massachusetts Avenue Cambridge, MA August 2018 We thank Pol Antras, Costas Arkolakis, Hanming Fang, David Dorn, Gordon Hanson, Ann Harrison, Amit Khandelwal, Fernando Parro, Daniel Xu, and seminar/conference participants at University of Pennsylvania, Johns Hopkins SAIS, MIT, George Washington University, UIBE, and Columbia University for helpful comments. The paper represents the personal views of the authors, and all errors are the responsibilities of the authors alone. The views expressed herein are those of the authors and do not necessarily reflect the views of the National Bureau of Economic Research. NBER working papers are circulated for discussion and comment purposes. They have not been peerreviewed or been subject to the review by the NBER Board of Directors that accompanies official NBER publications by Zhi Wang, Shang-Jin Wei, Xinding Yu, and Kunfu Zhu. All rights reserved. Short sections of text, not to exceed two paragraphs, may be quoted without explicit permission provided that full credit, including notice, is given to the source.

2 Re-examining the Effects of Trading with China on Local Labor Markets: A Supply Chain Perspective Zhi Wang, Shang-Jin Wei, Xinding Yu, and Kunfu Zhu NBER Working Paper No August 2018 JEL No. F16 ABSTRACT The United States imports intermediate inputs from China, helping downstream US firms to expand employment. Using a cross-regional reduced-form specification but differing from the existing literature, this paper (a) incorporates a supply chain perspective, (b) uses intermediate input imports rather than total imports in computing the downstream exposure, and (c) uses exporter-specific information to allocate imported inputs across US sectors. We find robust evidence that the total impact of trading with China is a positive boost to local employment and real wages. The most important factor is employment stimulation outside the manufacturing sector through the downstream channel. This overturns the received wisdom from the reduced-form literature and provides statistical support for a key mechanism hypothesized in general equilibrium spatial models. Zhi Wang Schar School of Policy and Government George Mason Universty 3351 Fairfax Drive, MS 3B1, Alington, VA zwang36@gmu.edu Shang-Jin Wei Graduate School of Business Columbia University Uris Hall Broadway New York, NY and NBER shangjin.wei@columbia.edu Xinding Yu School of International Trade and Economics University of International Business and Economics Beijing , CHINA yuxd@uibe.edu.cn Kunfu Zhu University of International Business and Economics Beijing , CHINA zhukunfu@163.com

3 1. Introduction Trade in intermediate goods has been growing steadily over time (Hummels, Ishii, and Yi, 2001; Johnson and Noguera, 2017; Koopman, Wang, and Wei, 2014). This can alter how imports affect the labor market of the importing countries. In this paper, we show that incorporating a supply chain perspective in a cross-regional reduced-form specification can overturn the received wisdom in the literature with a similar specification that looks only at the direct competition channel (Autor, Dorn, and Hanson, 2013). The paper also provides statistical support for a key mechanism hypothesized (but not tested) in recent general equilibrium spatial models used to assess the effects of trade shocks on labor markets. In 2000, US imports of intermediate goods from China was 14.8 billion US dollars, accounting for 28.6% of that year s total imports from China. This number increased to 63.2 billion USD in 2007, and doubled again to billion USD in 2014 (accounting for 37.5% of total imports from China that year). Across industries, those that experience a high rate of growth in total imports from China tend to see an even higher growth rate in intermediate inputs (see Figure 1). US firms that use these inputs potentially expand employment. Antras, Fort, and Tintelnot (2017) estimate that US manufacturing firms that source intermediate inputs from abroad tend to increase their production and may even buy more inputs from domestic manufacturing firms simultaneously. In this paper, we suggest that many nonmanufacturing firms also use imported inputs from China and increase their operational scale as a result. Indeed, the employment gains by non-manufacturing firms that can be traced to trading with China will be shown to dominate those of manufacturing firms. An influential paper by Autor, Dorn, and Hanson (2013), using a cross-regional reduced-form regression specification, shows that US regions with greater exposure to competition from China experience a greater decline in employment. Pierce and Schott (2016), another well-published paper that also focuses on the direct competition channel, reach the same conclusion that imports from China has reduced US manufacturing jobs and total employment. This paper adopts an explicit supply chain perspective, which is missing in Autor, Dorn, and Hanson (2013). We find that while the direct competition effect reduces manufacturing sector employment, an indirect upstream channel further exacerbates job losses in both manufacturing and non-manufacturing sectors. In other words, those US firms that do not compete directly with Chinese imports but sell their output to other US firms that are squeezed by Chinese imports also 1

4 experience job losses. However, the job gains from the downstream channel are not only statistically significant but also economically powerful enough to more than offset the combined negative effects from direct competition and the upstream channel. Once we account for all three channels of exposure to trading with China, the total effect for an average region is a net job increase of 1.27% (as a share of working age cohort) a year relative to a hypothetical region with no exposure to trading with China. We also find that 75% of the workers in an average region experience a real wage growth as a result of exposure to the China trade. To place this paper in the literature, it is useful to discuss four questions. First, how is it different from the previous attempt to incorporate a supply chain perspective in a cross-regional reduced-form specification? Second, how is it related to the emerging literature that uses a general equilibrium spatial model? Third, why would an indirect (positive) employment effect from a downstream channel be powerful enough to overwhelm a direct (negative) employment effect from import competition? Fourth, how to address possible endogeneity of US imports from China? Let us start with a comparison with the previous attempt to incorporate supply chain channels, namely Acemoglu et al. (2015). Methodologically, our paper differs from theirs in two ways. First, they do not separate intermediate goods from final goods in computing downstream exposure to China trade. Since the downstream effect is about how input costs are affected by imported Chinese inputs, that distinction is important. Second, Acemoglu et al. (2016) do not use exporter-specific information to allocate imports from China to the downstream sectors in the United States. In other words, they effectively assume that the imported inputs from China are allocated in the same way across US sectors as imported inputs from Germany or any other countries. Correcting these two items makes our approach better in line with a supply chain perspective. They turn out to make a big difference in the conclusion too. In particular, while Acemoglu et al (2016) reaffirm the conclusion of Autor, Dorn, and Hanson (2013) that trading with China causes a net job loss, we overturn this result. In addition, we find that the total real wage in the United States has been made higher by trading with China (which is not examined in their paper). We now relate our paper to the new literature on general equilibrium (GE) spatial models (Caliendo, Dvorkin, and Parro, 2018; and Adao, Arkolakis, and Esposito, 2018). Note that the cross-regional reduced-form specification used in this paper (as well as in Autor, Dorn, and Hanson, 2013) does not by itself speak about general equilibrium effects. Without information on interregional linkages, it would be misleading to extrapolate the results from the reduced-form 2

5 regressions to an aggregate effect on the labor market. GE spatial models, on the other hand, allow for inter-market linkages and can therefore speak meaningfully about the effects of trade shocks on aggregate labor market and on welfare. Our paper is a useful complement to the GE spatial models for four reasons. First, a key potential shortcoming of our specification is assessed to be unimportant quantitatively. In particular, both Caliendo et al., 2018, and Adao, et al., 2018, find that labor mobility across regions is very modest (with the median mobility across states being less than one half of one percent over a relatively long time period of ). (Autor, Dorn, Hanson, and Song, 2016, also report low inter-regional labor mobility.) In such a case, cross-regional reduced-form regressions are in principle valid. Since they are much easier to implement, it is useful to do them as a check on the performance of the GE models. Second, our paper provides a useful statistical test on a key mechanism hypothesized in both Caliendo et al. (2018) and Adao, et al. (2018). In particular, firms using imported inputs in both GE spatial models would expand their employment and much of the job expansion takes place in the service sectors (see, for example, Figures 1, 6, and 8 of Caliendo et al, 2018, and the associated discussions in the text). This mechanism is crucial for their conclusion that the overall effect of trading with China is an increase in the aggregate employment and aggregate welfare. However, neither paper provides a statistical test for the presence of such a mechanism. Indeed, the existing test in Acemoglu et al. (2016) appears to reject the significance of this mechanism. Our paper provides the first affirmative evidence that the downstream channel, especially outside the manufacturing sector, is statistically significant and economically powerful. In addition, our estimates could serve as useful moments for future GE models to match. Third, GE models make several assumptions that our paper does not. For example, Caliendo et al., 2018; and Adao, et al., 2018 assume that agents in their models have perfect foresight (in order for the models to be solved). Without having to make these assumptions, our results are potentially robust to alternative assumptions. Fourth, by using a specification that is essentially the same as Autor, Dorn, and Hanson (2013) except for the addition of two supply chain terms, our paper makes it transparent about what may be the crucial missing items in the existing reduced-form literature. In particular, stripped of multiple sources of complexity in GE models, our paper makes it easier to see that it is not the cross-regional mobility but rather downstream/upstream linkages (a particular form of cross-market linkages) that are responsible for the differences in the conclusions between the GE models of Caliendo et al. (2018) and Adao et al. (2018) and the reduced-form results of 3

6 Autor, Dorn, and Hanson (2013). The third comment is about how an indirect downstream effect on employment can be stronger than the direct competition effect. Intuitively, since a subset of the manufacturing sectors are responsible for most of the US imports from China, the direct competition channel only affects these sectors, which collectively constitute a small part of the US labor market. In comparison, the downstream channel benefits almost all sectors in the economy, including service sectors. Even research institutes, hospitals, schools, banks, law firms, government departments, and restaurants use imported Chinese made laptops, desktop computers, electric cables, communication devices, steel parts, tables and chairs, light bulbs, bed sheets, uniforms, or wash towels. This is true not only for the economy as a whole, but also for the vast majority of local labor markets. Note that the upstream channel also affects more sectors than the direct competition channel. Therefore, whether the downstream channel can ultimately overturn the existing results is an empirical question. The fourth comment is about possible endogenous nature of US imports from China. We employ three instrumental variable (IV) approaches to address this issue. The first is to use sector variations in imports from China by other high-income countries as instruments for the sector variations in US imports from China. This follows the spirit of the IV approach in ADH (2013). The second is to use the sector variations in the reduction of uncertainty after the United States granted permanent normal trading relations (NTR) status to China in 2000 to predict subsequent growth of US imports from China by sector. This follows the idea from Pierce and Schott (2017). The third IV approach combines the first two approaches, which allows us to perform an overidentification test for the validity of the instruments. We also investigate the effects of the China trade on real wages, another important labor market outcome that has also been studied by Autor, Dorn, and Hanson (2013), Ebenstein, Harrison, McMillan, and Phillips (2014), and Chetverikov, Larsen, and Palmer (2016). We show that the supply chain perspective makes a difference as well. In particular, if one focuses just on the direct competition effect, as Autor, Dorn, and Hanson (2013) and Chetverikov, Larsen, and Palmer (2016) do, one would find that workers in almost all initial income groups either experience a decline in real wage or no increase in real wage. In contrast, our supply chain perspective uncovers a different picture. For a region with an average exposure to trading with China relative to a region with no exposure, while there are winners and losers, the total effect of trading with 4

7 China is an increase in the real wage for 75% of American workers as well as an increase in the aggregate real wage. It might be useful to take note of the estimated impact of the China trade shock on employment in other countries. Using linked employer-employee data for (almost) the entire labor market in Denmark, Trailberman (2017) finds that trading with China does not result in a net increase in unemployment. Since the United States has a more flexible labor market than Denmark, it would seem surprising if the labor market outcome is indeed worse for the United States than for Denmark. We organize the rest of the paper in the following way. Section 2 provides some motivating facts about US-China trade. Section 3 introduces our empirical approach and data sources. Estimation results, model extensions and a set of robustness checks are presented in Sections 4-6, and Section 7 concludes. 2. Some Basic Facts about US Imports from China For US firms to benefit from imported inputs from China, one might conjecture that the Chinese imports are associated with cost savings for US firms. To check for plausibility of this channel, we examine unit import values at the HS 6-digit level. In the top graph in Figure 2, we present a bin scatter plot of changes in average unit value from 2000 to 2007 against increases in the share of China in US imports. More precisely, we divide all the increases in China s share in US imports at the HS 6-digit level into 20 equal-sized bins, and then plot the average value of the changes in the unit import values (on the vertical axis) for all observations in a bin against the midpoint of all changes in China s shares in the same bin (on the horizontal axis). The resulting bin scatter plot purges the noise from having too many data points on a raw scatter plot. We can clearly see a negative relationship between the two variables: those products for which China has become a relatively more important source tend to exhibit a greater decline in unit import values. As the supply chain perspective is about trade in intermediate goods, the bottom graph in Figure 2 presents a different bin scatter plot focusing on US imports of intermediate goods at the HS 6-digit level. We see the same pattern: Imported intermediate inputs become relatively cheaper when China becomes relatively more important as a source country. The data pattern can be confirmed more formally with the following regression model: ΔlnUnitPricei, =β0 +β1 ΔCHN-Share i, u i,

8 where i represents a 6-digit product under the HS classification system. ΔlnUnitPricei, is the change in log import unit price for product i averaged across all source countries from 2000 to 2007 (multiplied by 100). CHN-Sharei, is the change (in percentage points) in the quantity share (i.e., share in total weight or total physical units) of China in US imports of product i. We have done the regressions that give equal weights to all products as well as weighting each product in proportion to its total import values at the start of the sample period. They yield qualitatively similar results. In the first column of Table 1, we report the equal-weighted result in the first row, and the value-weighted results in the second row. In both cases, a negative coefficient means that an increase in the share of China in US imports and a decline in average unit value tend to go together. The effect is stronger when we weight the products according to their relative importance. Based on the coefficient in the first column, second row, an increase in the China share by one percentage point is approximately corresponding to a decline in the unit value by one percent. Note that the regression is done with all available observations (i.e., more than those in the bin scatter plots). To address endogeneity of the change in China s share in US imports, we use the change in China s share in other high-income countries imports as an instrumental variable 2. The IV results are reported in Column 2 of Table 1. We continue to find a negative and significant coefficient. Indeed, the point estimates are bigger than the corresponding OLS estimates. Based on 2SLS estimate in the value-weighted regression, an increase in the China share by one percentage point leads to a reduction in unit import price by 1.8%. In Columns 3 and 4, we re-do the regressions by focusing on intermediate inputs only. We find the slope estimates tend to be bigger than the corresponding estimates in the first two columns. Based on the 2SLS estimate from a value weighted regression (last column, last row), an increase in China s share in US imports of intermediate inputs by one percentage point tends to reduce the average US import price by 2.1%. Across all intermediate inputs, the median increase in China s share in US imports during was 8.23 percentage points; this translates to a reduction in import price by 17.4%. The top three most important intermediate inputs for the United States by import values are portable automatic data processing machines (i.e., laptops), transmission apparatus, and parts & accessories for data processing machinery (including computer parts), respectively. For these 2 Other high-income countries include Germany, France, Italy, Japan and the United Kingdom, hereinafter referred to as G5. 6

9 three intermediate inputs, the Chinese shares in US imports have increased by 67.6, 14.4, and 42.6, percentage points, respectively, during This leads to a much greater decline in import prices than the average or median across products 3. To summarize, the data on unit import values and China s shares in US imports are consistent with the notion that trading with China has generated substantial cost savings for US firms. It may also be useful to look at some macro facts regarding the relationship between US unemployment and the US trade deficit. Appendix Figure 1 plots the time series of US unemployment rate and US trade deficit as a share of total trade from 1960 to A striking feature that emerges from this graph is that the two variables tend to be negatively correlated: the US trade deficit tends to be large when the US labor market is strong (low unemployment) and small when the US labor market is weak. In other words, an increase in US net imports is unlikely to be associated with an increase in national unemployment. To zoom in on US trading with China, Appendix Figure 2 presents the time series of US unemployment rate and US trade deficit with China as a share of US total trade from 1990 to Again, the relationship is negative. The US tends to run a larger trade deficit with China when its employment situation is good and vice versa 4. While these macro facts are not a direct proof (as both are endogenous variables), they raise a question of whether trading with China systematically raises the US unemployment rate. Total employment is the sum of manufacturing and non-manufacturing employment. While the US manufacturing employment has been declining over the last two decades, the employment outside the manufacturing sector has been on a rise. While Autor, Dorn, and Hanson (2013) makes a case that the observed decline in US manufacturing employment is to a significant part due to trading with China, it implicitly assumes that the equally dramatic rise of non-manufacturing employment is not related to China. (Indeed, most service sectors are typically labeled as nontraded.) One way to interpret what we do in this paper is to discover and document that the rise in the non-manufacturing jobs is to a significant part also due to trading with China. 3 For big changes in the China share, the Jensen s inequality sets in, and the difference in log is no longer a good approximation for calculating percentage change in the import prices. 4 The same patterns are observed when we use US imports from China instead of US trade deficit with China. 7

10 3. Empirical Approach and Data Sources We now turn to the framework for examining the effect of trading with China on local employment in the United States. To maintain maximum comparability with Autor, Dorn, Hanson (2013) and Acemoglu et al. (2016), we intentionally keep the methodology and the data as close as possible to theirs. In particular, we use changes in employment and changes in exposure to trading with China at the Commuting Zone level as units of observation. We keep the differences at a minimum (by design), and they are introduced as a result of the supply chain perspective. First, we argument their specification by two additional terms capturing the downstream and upstream channels, respectively. Second, in computing the downstream exposure, we separate imported intermediate inputs from general imports. Third, we use a multicountry input-output table to capture the exporter-specific information on sector linkage. (This means that we do not have to assume that Chinese inputs are allocated across US sectors in the same way as German inputs or inputs from other countries.) These modifications make our measurement and framework more faithful to the spirit of a supply chain perspective. 3.1 Specification We run (variants of) the following regression: L Share Direct UP Down - k, t 0 1 k,t 2 k,t 3 k,t 4 k,0 k, t where k stands for one of the 722 Commuting Zones that cover the mainland US. The concept Commuting Zone was first developed by Tolbert and Sizer (1996), defined as an aggregation of counties that are characterized by strong internal commuting ties. This can be taken as a geographic area that constitutes a local labor market. It is the basic unit of observation in Autor, Dorn, and Hanson (2013). We estimate this model for two separate time periods t: a 7-year interval from 2000 to 2007 (t=2007), which is similar to Autor, Dorn, and Hanson (2013), and a 14-year interval from 2000 to 2014 (t=2014), which allows for more time for adjustment. To construct the dependent variable, we consider four mutually exclusive outcome variables, all measured as a share of the working age cohort in a Commuting Zone k: manufacturing employment, non-manufacturing employment, unemployment, and people not in the labor force. The four shares sum to 100%. L-Sharekt, is 8

11 100 times the annualized change in each outcome variable over the relevant time interval. Direct k,t, Upk,t and Down k,t are 100 times the annualized change in Direct, Upstream and Downstream exposures to trading with China in Commuting Zone k, respectively. They will be defined below in more detail. k,0 refers to a vector of control variables at the Commuting Zone level, including initial employment share in working-age population (age 16-64) and census divisions 5 fixed effects. 3.2 Three Channels of Exposure to the China Trade Shock To quantify the three channels of trade exposure (in terms of direct competition, downstream effect, and upstream effect, respectively), we start with a sector level measure and then convert them to commuting zone level measures based on each a sector s employment share in a region. The Direct Competition Channel The exposure to direct competition for Sector j is defined as the annualized change (in percentage point) in imports 6 from China of Sector j s products as a share of the sector s total absorption in year 2000: M M Direct =, t=2007 or 2014 C, U C, U 100 j,t j, 2000 j,t U * U U* t 2000 Y j, M j, E j, 2000 (1) where U Y j,t is total output of sector j in year t, M CU, j,t is US industry j s imports from China in year * U U* t, and M - E is US industry j s total imports from all sources minus exports to all destinations. j,t j,t The denominator Y +M - E equals total absorption of industry j at year This U * U U* j, 2000 j, 2000 j, 2000 definition of direct competition channel is identical to the Change in Import Penetration Ratio in Acemoglu et al. (2016). To control for US domestic demand factors in the US imports, we instrument the numerator in (1) with other high-income countries (G5) imports from China ( M CG, 5 j,t ) and replace the denominator by its 5-year lagged value as: 5 The United States Census Bureau divides the country into nine census divisions, including East North Central, East South Central, Middle Atlantic, Mountain, New England, Pacific, South Atlantic, West North Central and West South Central. 6 Trade values are converted to 2000 US dollars using the Personal Consumption Expenditure (PCE) deflator. 9

12 M M Direct =, t=2007 or 2014 C, G5 C, G5 IV 100 j,t j, 2000 j,t U * U U* t 2000 Y j, M j, E j, 1995 (2) We then convert the direct exposure to Chinese imports from the sector level to the Commuting Zone level based on the composition of the working age population in various sectors in each Commuting Zone. An Exposure to Direct Competition from China for Commuting Zone k from year 2000 to year t is defined as: L Direct = Direct, t=2007 or 2014 k,t,,2000 k j j Lk,2000 j,t (3) where subscript k indexes Commuting Zone, L k, j,2000 is the employment of industry j at Commuting Zone k in 2000, and L k,2000 represents total employment of Commuting Zone k in In other words, the commuting zone level measure of exposure to direct competition is the weighted average of the changes in import penetration ratios across sectors, with weights proportional to each sector s initial employment share. Following Autor, Dorn, and Hanson (2013), an instrumental variable version of the exposure to direct competition at the Commuting Zone level is defined as: L Direct IV = Direct IV, t=2007 or 2014 k,t,,1990 k j j Lk,1990 j,t (4) The weight on sector j s exposure to direct is the share of that sector in local employment in 1990 (the 10-year lag is proposed by Autor, Dorn, and Hanson, ). The Downstream Channel The Downstream Exposure for a Commuting Zone describes how it benefits from being able to use imported intermediate goods. We also construct it in two steps. First, at the sector level, it is a weighted average of all of its input g exposure to growth in China-sourced intermediate inputs (annualized to make it easy to compare across time periods): 7 When we use the labor shares in 2000, we obtain similar results (not reported to save space). 10

13 M -int M -int Down = w, t=2007 or 2014 (5) C, U C, U 100 Down g,t g, 2000 j,t g, j,2000 U * U U* t 2000 g Y-int g, M -int g, 2000 E-int g, 2000 The denominator is the total absorption of intermediate inputs at US industry j in year 2000, whereas the numerator is US imports of intermediates from China. As pointed out before, our measure focuses on imported intermediate inputs whereas Acemoglu et al. (2016) use all imports including final goods. The sectoral weights are proportional to each input sector s imports of intermediate goods from China: w Z (6) CU, Down g, j,2000 g, j,2000 CU, Zi, j,2000 i The numerator in the weight represents imports of intermediate input in sector g from China by US industry j in 2000, whereas the denominator is all intermediate inputs from China used by US industries j. Importantly, it does not assume that the Chinese and German intermediate inputs are allocated in the same way across US sectors (because we use an inter-country input-output table). In comparison, Acemoglu et al. (2016) and Feenstra et al. (2017) effectively make this assumption, and this assumption is rejected by the Inter-Country Input-Output Table. The downstream exposure at a Commuting Zone level is the weighted average of the sector level downstream exposure, with the weights proportional to each sector s employment share in 2000: L Down = Down, t=2007 or 2014 k,t,,2000 k j j Lk,2000 j,t (7) The instrumented version of the Downstream Exposure at the sector level is constructed as: M -int M -int Down = w, t=2007 or 2014 (8) C, G5 C, G5 IV 100 Down g,t g, 2000 j,t g, j,2000 U * U U* t 2000 g Y-int g, M -int g, 1995 E-int g, 1995 The instrumented version of the Downstream Exposure at a Commuting Zone level is the weighted average of the corresponding sector level measure, with employment share in 1990 as the sector weight. That is, the instrumented measure of Downstream Exposure for Commuting 11

14 Zone k is: L, =2007 or 2014 (9) IV k, j,1990 IV Down k,t = Down j,t t j Lk,1990 The Upstream Channel The Upstream Exposure captures how a Commuting Zone may be affected indirectly when their firms are at an upstream to those US firms that compete with Chinese imports directly. For a given sector j, the Upstream Exposure is the annualized change in sales-weighted average of the direct competition exposure of all of its customers: UP where weight w j, g, t is computed as: UP UP j,t = wj, g,2000 Direct g,t, t=2007 or 2014 (10) g w Z (11) UU, UP jg,,2000 jg,,2000 UU, Z ji,,2000 i UU where ji,,2000 Z represents US industry j s output sales to US industry i as the latter s intermediate UP input. Thus, the economic meaning of such a weight wjg,,2000 is the relative importance of US industry g for industry j as a percent of industry j s total sales in year The higher the percentage, the larger the impact of sector g s direct exposure to the China trade shock. We convert the sector level measure of upstream exposure to Commuting Zone level by making use of each sector s initial share in local employment: UP = k,t L,,2000 k j j Lk,2000 UP j,t, t=2007 or 2014 (12) The corresponding instrumental variable version is: L UP IV = UP k,t,,1990 k j j Lk,1990 IV j,t, t=2007 or 2014 (13) UP w Direct (14) IV UP IV where j,t j, g,2000 g,t g 12

15 3.3 Data Sources and Basic Statistics The construction of the downstream and upstream exposures requires the use of inter-country input-output (ICIO) tables. We use ICIO tables from OECD, which cover 64 countries and 34 industries from 1995 to The structure of the ICIO Table is presented in Appendix Table 1. The local employment data are derived from the U.S. Census microdata (5% sample for the year 1990 and 2000) and American Community Survey (ACS) microdata (for the year 2001 to 2014) provided by the IPUMS-USA database (Ruggles et al., 2015). These two datasets use a 5- digit numeric variable (PUMA code) to identify the Public Use Microdata Area where the respondent is located. The PUMA code is state-dependent, which must be read in conjunction with the 2-digit State FIPS code. We merge the Public Use Microdata Areas to 722 Commuting Zones by using the concordance between the 1990/2000 Census Public Use Micro Areas and 1990 Commuting Zones provided by Autor, Dorn, and Hanson (2013), and a crosswalk between the 2010 and 2000 version of Public Use Microdata Areas provided by the IPUMS-USA database. Both census and ACS data provide information on a respondent s employment status: whether he/she is in the labor force, currently unemployed, and in which industry is the employment. The respondents' wage income, gender and educational attainment are also available. Table 2 shows the three channels of exposure to China trade at the sector level in terms of annualized percentage changes and exposure to exports to China. Taking Direct Exposure to imports from China (ΔDirect) as an example, the mean of represents an annual increase of 0.224% on average during 2000 to In Figure 3, we plot the three channels of exposures at the industry level during for all 34 industries in the OECD ICIO database. The direct competition channel only affects the manufacturing sectors in which China has comparative advantage or runs a large trade surplus from processing and assembling trade. Those manufacturing industries, such as Textile Products, Computer and Electronic Products and Electrical Equipment, account for a large portion of imports from China, but collectively only account for a small part of the US labor market. While the upstream exposure is more important in the manufacturing sector than service and primary sectors, the downstream exposure benefits almost all sectors in the economy. We now turn to commuting zone level measures. For all 722 Commuting Zones, as shown in 13

16 Table 3, the exposure to direct competition has increased by an average of 0.112% a year during and 0.082% a year during The top five Commuting Zones that have experienced the most direct competition during are Rome, GA; Hickory, NC; Morganton, NC; Martinsville, VA and Talladega, AL, respectively. The five Commuting Zones that exhibit the least exposure to direct competition are Gunnison CO; Granby CO; Winnemucca NV; Elko NV and Reno NV, respectively. Interestingly, both the indirect upstream and downstream exposures also increased during the two periods. Similar to the industry level results, the increase in the downstream exposure is greater than those of the other two channels, partly because the imports of intermediate goods from China has grown faster than the imports of final goods. As shown in Table 4, there has been a trend decline in the manufacturing employment share. In comparison, non-manufacturing jobs exhibit a steady increase (at the rate of 0.231% a year during , offsetting the loss of manufacturing jobs, which was 0.23% a year). During , the labor force non-participation rate decreased at the rate of 0.048% a year, slightly more than offsetting an increase in the unemployment rate of 0.047% a year. In comparison, over the 14-year span of , both the unemployment rate (out of the local working age cohort) and the labor force non-participation rate went up. This is likely due to the massive job destruction caused by the Subprime Mortgage Crisis and global financial crisis during rather than trading with China. 4. Estimation Results 4.1 An Incompletely-Specified Model that Only Looks at the Direct Competition Channel This sub-section follows the specification in ADH (2013) by looking only at the direct competition channel, ignoring the downstream and upstream channels. This is to ensure that we can produce the same results as they do when the model specification is the same. The regression results are shown in Table 5. On the impact on US manufacture employment, our estimation results are consistent with Autor, Dorn, and Hanson (2013) and Pierce and Schott (2017). Both the OLS and the 2SLS estimates indicate a negative impact on US manufacturing employment. Using the 2SLS estimates as an example, a one percentage point rise in direct exposure to Chinese imports will reduce manufacturing employment by 4.2 percentage points per year from 2000 to 2007, and

17 percentage points per year from 2000 to These numbers are comparable to ADH (2013). (The point estimates are slightly larger than theirs because we scale our dependent variables by the working age cohort, whereas they scale everything by the labor force.) The results in Table 5 suggest that the slight difference in the definitions of the dependent variables does not make any qualitative difference. In column 2 of Table 5, we report the results on non-manufacturing employment share. The same exposure to direct competition raises employment in the non-manufacturing sector (e.g., laidoff steel workers may be re-employed as restaurant wait staff) but the increase is smaller than the decline of manufacture employment, resulting in a negative effect on total employment (Column 5). In Columns 3 and 4, we see that both the unemployment rate and the not-in-the-labor-force rate go up in response to the exposure to direct competition from the Chinese imports. 4.2 Accounting for Supply Chain Channels We now introduce the Upstream and Downstream exposures to the regression specification. The benchmark results are reported in Table 6, with the first stage regressions shown in Table 7. While the results for the two time periods are similar qualitatively, we use the results for the period to illustrate the interpretation of the results. The direct competition effect is negative on manufacturing employment (a decline with an elasticity of -3.5% in Column (1) of Table 6). This number is smaller than the corresponding number (-4.2) in Column 1 of Table 5 without the supply chain variables. The direct effect on non-manufacturing jobs is positive (Column 2), reflecting possibly laidoff manufacturing workers (from both a direct competition effect and an indirect upstream effect) working in service sector jobs such as restaurant servers. The direct effect leads to fewer people staying outside the labor force (Column 3), and the impact on officially recorded unemployment is small and not statistically significant (Column 4). The supply chain perspective produces two terms with opposite signs. On the one hand, adding the upstream effect augments the negative impact on the US labor market. This is especially true for service sectors jobs that provide inputs to those manufacturing firms that compete with China imports directly (Column 2). On the other hand, the downstream channel produces a job gain, especially in the non-manufacturing sector (with an elasticity of 5.6%). The downstream channel also raises the labor force participation rate. 15

18 Putting the results from Columns 1 and 2 together, we see that the downstream channel produces a net gain in jobs with an annualized elasticity of 5.9% during and 3.6% during , respectively (column 5). Such positive effects are significant during both the and periods, but particularly large during the first period. It is noteworthy that the F statistics from the first stage for three endogenous variables are 298.9, and for the regression, and 127.4, 69.8 and for the regression, respectively (Table 7). They allow for an easy rejection of the weak IV null hypothesis. While the values of the F statistic are larger than many applications of the 2SLS technique, there is no theoretical upper bound for the statistics. To interpret the results and translate the estimates into economic magnitude, let us consider a hypothetical average commuting zone whose three channels of exposure to trading with China are equal to the average values across all commuting zones, as reported in the left panel of Table 3, and compare it to another hypothetical commuting zone that has no exposure to trading with China. We can convert these estimates of the elasticities in Table 6 to estimates of the job market responses by combining with the mean values of the regressors reported in Table 3. The implied labor market reactions are reported in Table 8. The effect of the exposure to direct competition in the average commuting zone is a job loss in the manufacturing sector at the rate of 0.39% a year (Column 1 of Table 8). Incorporating the upstream effect would raise the negative effect on manufacturing jobs to 0.63% a year (-0.39% %= -0.63%). On the other hand, the sum of the direct and upstream effects also produces a loss of nonmanufacturing sector jobs at the rate of 1.34% a year (0.87% %= -1.34%). (Those service firms that used to provide inputs to the directly affected manufacturing firms also shed jobs.) The sum of the direct competition and upstream exposure produces a reduction in total employment (0.48% %= -1.98%, Column 5). This reduction in total employment can be decomposed into some decrease in the labor force participation rate (Column 3) and some increase in the reported unemployment (Column 4). However, this is not the whole story. In particular, the downstream channel produces large job gains in the non-manufacturing sector (at the rate of 3.08% a year, Column 2) and a small increase in jobs in the manufacturing sector (at the rate of 0.16% a year, Column 1). 16

19 When we sum up all three channels (downstream, upstream, as well as the direct competition effects) in both manufacturing and non-manufacturing sectors, the total effect of trading with China is a net job gain of 1.27% a year during (and a job gain of 0.69% a year during ) as reported in Column 5. Of course, many factors affect job market performance including technology and regulations besides trade. What the estimates in Table 8 suggest is that these other factors may well have led to job losses, but trading with China has more than mitigated the job loss. Another way to provide economic interpretations to the estimation results is to compare two commuting zones whose exposure in terms of the direct competition channel is at the 25 th and 75 th percentiles of the entire distribution, respectively. To be concrete, the city of Plainview in Taxes - at the 25 th percentile of the distribution - experienced a relatively small exposure to direct competition from Chinese imports during In comparison, the city of Douglas in Illinois - at the 75 th percentile of the distribution - experienced a relatively large direct competition effect from Chinese imports. Unsurprisingly, by our estimation, Douglas loses more manufacturing jobs than Plainview due to competition with Chinese imports. Once we have picked this pair of cities, their indirect exposure to Chinese imports in terms of the downstream and upstream exposures can also be calculated. We summarize the relative effects of trading with China on the job markets in these two commuting zones in Table 9. First, if we use an incomplete specification that only looks at the direct competition channel (i.e., using the same specification as Autor, Dorn, and Hanson, 2013), we would have concluded that, relative to Plainview, Douglas has experienced an additional loss of manufacturing jobs at the rate of 0.22 percentage points a year, and an additional loss of total jobs at the rate of 0.15 percentage points a year. In other words, greater exposure to direct competition with China produces a greater relative job loss. Second, when we use a more complete specification that takes into account the supply chain channels, we will find the opposite result. Even though Douglas suffered more from a combination of a direct competition effect and an indirect upstream effect in the manufacturing sector, this is completely offset by job expansion in the non-manufacturing sector 8. In fact, taking into account 8 In this example, because the downstream exposure is big in both Plainview and Douglas, the difference in their downstream exposure is relatively small. 17

20 all three channels of exposure to Chinese trade, Douglas has a small net job gain of 0.01% a year relative to Plainview. Another way to see how the supply chain perspective alter the inference is to examine the commuting zones most negatively hit by a direct competition effect. An important feature to note is that in almost all commuting zones, non-manufacturing employment tends to dominate manufacturing employment. (At the commuting zone level, there are no single-factory towns.) For example, in Detroit in 2000, while 15% of the 790,000 people in the age cohort are employed in the manufacturing sector, 53% are employed outside manufacturing. (5.4% are unemployed, and 29% are not in the labor force). As we noted earlier, most sectors, including those that are sometime labeled as non-tradable sectors, can in fact benefit from being able to use imported intermediate inputs from China. In terms of the negative job effects from the direct competition channel, the Detroit Commuting Zone is not the worst hit in the country. The five Commuting Zones that are most severely affected in terms of loss of manufacturing jobs via a direct competition channel are: Rome, GA; Hickory, NC; Morganton, NC; Martinsville, VA and Talladega, AL, respectively. Table 10 reports the estimated manufacturing job loss from the direct competition channel in these five places (relative to a hypothetical region with no exposure to Chinese imports). By construction, they all show a large negative job effect in the manufacturing sector. Importantly, the table also reports the downstream and upstream effects in both the manufacturing and non-manufacturing sectors. (The calculations are done when each is compared to a hypothetical commuting zone that is unaffected by trading with China in any way.) An important take-away message is that taking into account the supply chain channels is important, and the job expansion effect in the non-manufacturing sector (that can be traced to trading with China) is economically powerful enough to offset any job loss in the manufacturing sector. In the end, the total effect of trading with China does not produce a net job loss in any of these five commuting zones. It may be instructive to compare the total employment effect and the direct competition effect across all commuting zones through two graphs. In Figure 4, we plot the actual employment change against the direct exposure to imports from China across the 722 Commuting Zones from 2000 to We can see a negative relationship between the two: on average, those Commuting 18

21 Zones that experience greater growth in imports from China tend to experience a greater decline in local employment. This of course is a graphic representation of the ADH result. In Figure 5, we plot the total employment effect after taking into account all three channels of exposure to the China trade. Strikingly, the relationship between the total employment change and the total China effect across all Commuting Zones is positive when the supply chain perspective is incorporated. In other words, those regions with greater exposure to total China effect tend to experience a relatively greater increase in local employment. Basically, non-manufacturing industries are a bigger part of a local labor market than manufacturing industries in all commuting zones, and the expansion of local non-manufacturing jobs can be systematically and statistically traced to the ability of the United States to import China made intermediate inputs. Note that in the absence of information on cross-regional mobility, one cannot extrapolate the relative-relative results from such reduced-form regressions to the aggregate effect in local labor markets. However, GE spatial models of Caliendo et al. (2018) and Adao et al. (2018) have found the inter-regional mobility to be low. Adao et al. (2018) explicitly conclude that the results from the reduced-form regressions are in principle valid. Since US employment tends to be stronger when US imports more from China or when it runs a larger trade deficit with China (Appendix Figure 2), it would seem easier to reconcile the aggregate employment patterns with our conclusion than with that of Autor, Dorn, and Hanson (2013). 5. Extensions and Robustness Checks 5.1 Alternative Measures of Downstream and Upstream Exposures Our benchmark measures of upstream and downstream exposures keep the diagonal elements in the input-output matrix in computing the weights. There are two potential issues. First, since these elements are reflected in both the direct competition channel and the two indirect value chain channels, there is some double counting in these measures. Second, the double counting makes it more likely that the indirect channels and direct channel are collinear. Pairwise correlations among the three measures are presented in Table 11; the multicollinearity problem appears most serious between the direct competition channel and the upstream channel. This makes it hard for the regression coefficients to be estimated precisely. 19

22 As a robustness check, we compute an alternative pair of downstream and upstream measures that exclude the diagonal elements in the input-output matrix. We re-do the regressions in Tables 6-8 with the new set of regressors, and report the corresponding results in Tables and Appendix Table 2. As it turns out, our key results and interpretations are not affected. In particular, while a direct competition effect (and an upstream effect) produces a job loss, this is more than offset by a job expansion effect from a downstream channel. Overall, trading with China does not produce a net job loss once the supply chain channels are taken into account. 5.2 Alternative Instrumental Variables As the second set of instrumental variables to control for possible endogenous nature of imports from China, we use differential reductions in uncertainty facing imports from China across products with US granting Permanent Normal Trade Relations (NTR Gap) to China in 2000 to construct additional instrumental variables for the increase of imports from China. This follows the idea in Pierce and Schott (2016). We take five steps to calculate the NTR Gap for each industry (at the level of OECD ICIO industry). First, we aggregate the Column 1 (MFN or NTR) tariff rates that the United States offers to WTO members and the Column 2 (non-ntr) tariff rates to 6-digit HS level from the original 8-digit HS level provided by Feenstra, Romalis and Schott (2002). Second, using the concordance provided by OECD, we match the 6-digit HS code to OECD ICIO industry via ISIC revision 3 code. Third, the NTR gap for each OECD ICIO industry is calculated as the percentage difference between the NTR and non-ntr tariff rates. 9 Fourth, we calculate the instrumented versions of the upstream and downstream exposures with NTR gaps by using equation (7) and (10). It is worth noting that the uncertainty removed by the NTR affects both final consumption goods and intermediate goods. Based on our definition of downstream exposure, only the latter has a cost reduction effect on downstream sectors. We use the BEC classification to separate imports of intermediate goods and those of final goods, and calculate separately the NTR Gaps for consumption goods and intermediate goods. Finally, in the fifth step, we convert the industry level NTR gaps to the Commuting Zone level based on each 1+non-NTR tariff rate 9 The NTR GAP = 1. Another way to calculate the NTR Gaps is to directly use the difference between 1+ NTR tariff rate the US column 2 and MFN tariff rates. Our baseline approach recognizes that a reduction in the tariff rate from 5% to 1% is a proportionately bigger reduction than a reduction from 25% to 21%. The two approaches are qualitatively similar. 20